Skip to content

Glossary

The terms these pages use, each with the page that explains it.

Candidate term
A class or property a unit minted that the document's ontology plan did not list. Consolidation decides it; without a decision it is dropped. See How OntoCast works.
Catalog
The set of ontologies OntoCast knows: those loaded at startup, uploaded to /ontologies, or created by earlier runs. Every ontology context is taken from it. See Ontologies and facts.
Completion pass (gap pass)
A small call after the critic, made only for a part with a gap: a measurement the render missed, or an observation that names nothing it is about. It only adds facts and links. See How OntoCast works.
Consolidation
The step that decides every term a document minted before it is written: keep, merge into another term, generalise, demote to a property value, or drop (ONTOLOGY_CONSOLIDATION). See How OntoCast works.
Content unit
One part of a document, cut along its sections, that is extracted on its own and in parallel with the others. See How OntoCast works.
Coreference resolution
At most one LLM call per document, after the merge, that decides the entity pairs the merge could not settle (AGG_LLM_COREFERENCE). See How OntoCast works.
Critic
The LLM call that reviews a render and names the statements to fix. OntoCast applies the fixes as a patch without another call. See Validation layers.
Curation
A pass over a catalog ontology, across every document processed so far, that keeps, merges, generalises, demotes or drops the terms documents added (ontocast curate). See Curating the catalog.
Entity disambiguation
Deciding which entities from different content units are the same thing, and merging them. See Entity disambiguation.
Facts
The graph of what a document states: the entities it mentions and their relations and values, in the terms of the ontology. See Ontologies and facts.
Facts namespace (cd:)
https://growgraph.dev/facts/, where every entity extracted from text is named. The ontology's own namespaces are never extended by facts. See Ontologies and facts.
GraphUpdate (patch)
What the model returns instead of a whole graph: the triples to insert and the triples to delete. See Ontologies and facts.
Head
The version of an ontology no other version names as prior. A run reads one head per ontology; an ontology with several heads is diverged, and runs cannot extend it until an upload (PUT) merges it. See Ontologies and facts.
Extension ontology
The ontology CURRENT_DOMAIN/<id>-ext-<hash> that holds the terms OntoCast adds to an external ontology, which it imports. <id> is the external's id and <hash> is taken from its IRI. See Ontologies and facts.
External ontology
A catalog ontology whose IRI is not under CURRENT_DOMAIN, such as a published vocabulary. OntoCast reads it and never writes to it.
Home ontology
Where a document's new vocabulary goes: an existing catalog ontology it extends, or a new one minted under CURRENT_DOMAIN, resolved once per document. See Ontologies and facts.
IRI
The web-style name of anything in a graph, such as https://example.org/onto#Device. Two things with the same IRI are the same thing, wherever they appear.
Label
The readable name of a term or entity (rdfs:label). The IRI identifies it; the label is what a person reads.
Ontology
A graph of classes, properties and declared individuals, identified by its IRI and stored in versions. See Ontologies and facts.
Ontology context
The slice of the catalog a content unit is shown: one chosen ontology, the terms vector retrieval finds, or one fixed ontology (ONTOLOGY_CONTEXT_MODE). See Choosing ontology context.
Ontology plan
The classes and properties a document adds to its domain, listed by one LLM call over the whole document before its units render and checked against the catalog without a model (ONTOLOGY_PLAN_ENABLED). Units mint from it, and the terms it reuses are shown to the facts units. See How OntoCast works.
Prefix
A short alias for the start of an IRI, so ex:Device can stand for https://example.org/onto#Device in Turtle. It changes how a file reads, not what it means.
Provenance
The link from each fact to the content unit it came from, written with RDF 1.2 reifiers. See Ontologies and facts.
RDF
The W3C standard data model for graphs: a graph is a set of triples. OWL and SHACL are written in it, and Turtle is one way to write it down.
Render
The LLM call that extracts an ontology patch or a unit's facts. See How OntoCast works.
Render mode
Which blocks a run executes: ontology_and_facts, ontology or facts (RENDER_MODE). See How OntoCast works.
Run manifest
The JSON file ontocast process writes beside each facts file (<name>.run.json): the model, the settings and what the run cost. See Reading run telemetry.
Route
What happens to a content unit before extraction: extracted, extracted as bibliographic records, kept but flagged as non-content (front or back matter, a table of contents), or dropped, with the reason recorded. See Structured documents.
Route class
What a section label's sections hold, set in its section schema: content, boilerplate, non_content (front and back matter) or citations (a reference list). Evidence for the route, never a decision on its own.
Section label
The section a content unit came from, such as methods or results, used to select what to extract. A label is strong when it comes from an exact heading, a heading placed inside a strong section by its number or level, or a reference list or acknowledgements recognised from the content, and weak otherwise (a keyword in a longer heading, a heading's nearest label by embedding, inheritance by wording or from a weak parent, front matter recognised by position, other guesses from the content, the model); only a strong label can take part in dropping a unit. Each step's vote is kept as the label's evidence. See Structured documents.
Section schema
The set of section labels for one kind of document, such as academic or financial. See Structured documents.
SHACL shape
A constraint the extracted facts are validated against, loaded from FACTS_SHAPES_DIR or uploaded to /shapes. See Validation and SHACL.
Tenant and project
Request parameters that partition the triple store and the vector store, so that several users or projects share one server without seeing each other's graphs. See Tenancy.
Term sheet
A compact form of the ontology context in facts prompts: one line per term instead of a serialized graph (ONTOLOGY_CHAPTER_FORMAT). See Performance tuning.
Triple
One statement in three parts: subject, predicate, object. In ex:Device_C ex:madeOf ex:Perovskite . the subject and object are entities and the predicate is the link between them. The object can also be a value, such as a number or text.
Turtle
A compact text format for RDF: triples ending in ., with prefixes keeping IRIs short. OntoCast returns graphs in it, and its models write it by default (LLM_GRAPH_FORMAT).
Unit pipeline
The per-unit loops run on a single piece of text, with no conversion, chunking or merging: POST /process_unit or run_unit_pipeline. See Embedding OntoCast in your agent.
User instruction
Your own guidance, added to the prompts of one request (facts_user_instruction, ontology_user_instruction, ontology_selection_user_instruction). See Writing user instructions.
Vector retrieval
Choosing a unit's ontology context by searching an index of the whole catalog (LanceDB or Qdrant) for the terms the unit needs. See Retrieval.
Version IRI
The named graph one immutable version of an ontology is stored as, <iri>#<key>; its header names the prior version(s) it was made from. See Ontologies and facts.